Latest / Daily AI, by AI / Frontier AI Pricing Power Collapses as Grok and DeepSeek Match Top Models
Transcript
- Good to be with you. It’s just me on the mic today, and frankly, the pace of these releases is demanding our full attention. There is quite a bit to unpack, and I want to ensure you understand the strategic implications behind what happened over the last twenty-four hours. Following the SpaceX-Cursor consolidation we covered back on August 11th, the fruits of that union have arrived faster than many expected. SpaceXAI officially released Grok 4.6... a model that matches frontier performance against Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol while charging sixty percent less. And in typical fashion, Elon Musk has already teased that Grok 4.7 is just weeks away. DeepSeek also quietly dropped V4-Pro-0813 overnight. As we’ve been tracking since their geopolitical constraints first made headlines on January 15th, this team continues to defy the compute-scarcity narrative. They are pricing output at just eighty-seven cents per million tokens—less than two percent of the cost of Fable 5—while matching it nearly point-for-point on tool benchmarks. In other developments, Google DeepMind launched SL2T. This is a sign-language-to-text model built into Gboard and Live Transcribe on the Pixel 11, allowing Deaf users to sign ASL directly into any text field in real time. On the financial front, Lovable raised four hundred million dollars at a thirteen point three billion dollar valuation. That valuation has more than doubled since December, with annualized revenue approaching six hundred million by the end of this month. Finally, researchers at Germany's University of Tübingen have found a method to extract the hidden reasoning traces that frontier models generate mid-task. They are pulling these from OpenAI, Anthropic, and Google models via API—and in some cases, they are surfacing buried passwords and API keys. Here is the strategic headline: we are witnessing the death of the "Impossible Triangle." For years, the AI industry operated on an unspoken rule. A model could be smart, or cheap, or fast—but humans had to pick two. That constraint was the frontier labs' greatest business asset. It allowed OpenAI to charge thirty dollars per million output tokens for Sol, and Anthropic to justify fifty dollars for Fable 5. The premium existed because the gap was real. If an enterprise needed frontier intelligence, they paid frontier prices and waited for frontier-length processing times. In the span of two hours yesterday, two separate releases shattered that logic. DeepSeek shipped V4-Pro and SpaceXAI shipped Grok 4.6. Neither coordinated, but together, they have effectively killed the alibi for high-cost inference. Thom isn’t joining me on the mic today, so I’ll channel his technical energy for a moment to look at the benchmarks... he’d be the first to point out that these efficiency gains are the real story. Grok 4.6 scores a 61 on Artificial Analysis' Intelligence Index—passing GPT-5.6 Sol and sitting just behind Fable 5 at 62. On agentic benchmarks, it’s particularly striking. Tasks that drag past thirty minutes on Sol finished in under twenty on Grok. On the AA-Briefcase benchmark, Grok reached Fable 5-tier performance while averaging fifty-three turns—compared to one hundred and three turns for Claude Opus 5 Max. Hmm. That is not a marginal gain. That is half the compute consumption for comparable output. DeepSeek's V4-Pro-0813 took a different angle, matching Fable 5 on Terminal Bench and outcompeting Opus 4.8 on Cybergym and AutomationBench. The architectural insight both models share is an obsession with agentic efficiency: fewer turns, tighter context, and less token waste. This matters because the cost of running an agent isn't just the per-token price—it’s turns multiplied by tokens multiplied by time. Compress all three, and the industry enters a new phase. The pricing gap here is structural. Grok 4.6 comes in at two dollars per million input tokens and six dollars per million output tokens. Compare that to Fable 5 at fifty dollars. DeepSeek is even lower at eighty-seven cents. At those numbers, an executive team running agents all day isn't facing a marginal cost decision—it’s a category decision. The question becomes: are your teams paying frontier prices for a shrinking frontier advantage? The answer is increasingly no. The Ramp AI Index, published this month, flagged slowing adoption of OpenAI and Anthropic among businesses, with alternative models capturing more of the incremental spend. Lovable's thirteen billion dollar valuation tells the other half of the story... the application layer is exploding precisely because inference costs are collapsing. When running an agent gets cheap enough, developers build more agents. Cognition reportedly hitting one billion in ARR underscores this. Cheap inference doesn't hurt the AI economy—it accelerates it at the application layer while compressing margins for the labs. The labs most exposed here are the ones whose business model depends on the premium holding. OpenAI and Anthropic have invested enormously in differentiated capability. If that differentiation narrows from a canyon to a crack, pricing power follows. Twelve months ago, describing Grok as a frontier competitor would have been a punchline. Now, the SpaceX plus Cursor combination is being cited as the number three frontier lab. That is a remarkable shift for 2026. What changed? The Cursor partnership gave xAI a direct feedback loop with professional developers doing real agentic work. That loop is exactly the kind of signal that sharpens a model faster than benchmark chasing. Meanwhile, DeepSeek's releases demonstrate that frontier capability is increasingly decoupled from massive compute budgets. The pressure this creates isn't just on pricing—it’s on cadence. Musk has telegraphed that 4.7 arrives in less than a month, post-trained on SpaceX's own engineering data. If that specialization story holds, xAI isn't just competing on intelligence scores—it’s competing on domain expertise baked into the model itself. That is a much harder moat to erode. For the established labs, the strategic response options are narrowing: accelerate, cut prices, or find differentiation that benchmarks can't capture. There is a behavioral shift embedded in this story as well. When frontier intelligence gets cheap enough, the question stops being "can we afford to run this?" and starts being "what should we point it at?" OpenAI's own enterprise usage studies show high-usage firms generating eight times more output tokens per user than typical enterprises. Those humans aren't using AI more because it got smarter—they’re using it more because the cost-benefit math changed. The user experience story is illustrative here. The interfaces that stick aren't always the most powerful—they are the ones that feel like messaging a teammate. Simplicity and presence beat raw capability for sustained daily use. As models converge on quality, the interface layer becomes the primary battleground for human behavior. The labs that figure out daily ritual adoption will capture the moat that pricing wars cannot buy. Here is the action plan for executives this week. First, run a cost audit on your current model spend with agentic workloads specifically in scope. If your developers are running multi-hour tasks on Fable 5 or Sol, benchmark Grok 4.6 on those same tasks. The efficiency gap on long-horizon work is where the real savings live. Second, treat DeepSeek V4-Pro as a serious option for internal tooling, with appropriate caveats. The pricing is extraordinary, but the risks—potential price hikes and data governance questions—are real. Build a tiered model strategy: frontier models for client-facing work, and cost-optimized alternatives for internal automation. Finally, if you are building on top of AI models, the Lovable and Cognition funding rounds tell you where the market is going. The application layer is where value is accruing as inference commoditizes. The key takeaway: if your competitive advantage lives primarily in access to a specific model, that advantage is eroding. Build the workflow, the feedback loop, and the domain expertise that survives a model swap—because those swaps are coming faster than anyone predicted even a few months ago.